Lecture 21 - Game Theory, Strategic

Equilibrium, and Repeated Interactions

14.03 Spring 2003

1 Game Theory, Strategic Equilibrium and Re-

peated Interactions
• The study of the economics of information gives rise to several important
modeling constructs that were not modeled in consumer choice:

1. Asymmetric information
2. Conjectures about the behaviors of others
3. Interdependence of your actions on the conjectured choices of others
— leads to strategic behavior.

• These attributes required us to specify the features of the environment

more carefully:

— Information/uncertainty
— Conjectured behavior
— Interdependence

• Naturally gives rise to game theory: a tool for analyzing strategic inter-
actions among individuals (people, firms) in an economic setting.
• Like all previous models, we have rational actors maximizing their well
being in a well specified environment.
• What is new here is small N , that is a small number of actors such that the
best choices for any one actor depends intimately on the choice of a small
number of other actors. This is quite different from a standard market
setting where everyone is a “price taker.” In this sense, the situations
described by game theory resemble a multi-player game with strategies,
payoffs and concealed information.
• An economic game has three elements:

1
 1. Players
2. Strategies
3. Payoffs (utilities, states)

• There are both cooperative and non-cooperative games. In cooperative

games, players can make binding commitments. In non-cooperative games
(which are much more interesting), players cannot make binding commit-
ments. These means that the only actions they can credibly ‘commit’ to
are self-interested actions.
• Players

— 2 persons/firms
— N persons/firms
— Can also be against nature. In that case, nature moves stochastically.

• Strategies

— Each course of action that is open to a player is a strategy.

— Usually this is a finite set; often limited to 2.
— In a non-cooperative game, players cannot make binding commit-
ments about which strategies they will play.

• Payoffs

— Final returns to players at conclusion of game.

— Usually measured in utility or profits.
— Utilities not necessarily assumed to be the same across players.
— Payoffs assumed to include all aspects of the pleasure/pain of final
outcome. If for example you take extra glee in demolishing your
opponent, this will be incorporated in final payoff and hence is not
an additional consideration beyond that payoff.

• Two ways to write down a game:

— Extensive form: Tree with nodes

— Normal form: Payoff matrices

• Depending on the game, one or the other notation will typically be more
useful.

2
1.1 Example: Dormitory game
• Player A prefers to play music quietly, but dislikes hearing loud music
from the adjoining room.
• Player B is a head-banger. The more noise the better.
• This is a simultaneous move game.
• See Figure 21#1
Dormitory game:
Extensive form 21#1a
A

S = soft
S L
L = loud

B B

L S L
S

(6,3) (2,4) (5,4) (3,5)

(A,B) (A,B) (A,B) (A,B)

Dormitory game:
21#1b
Normal form
B

S L

S 6,3 2,4
A

L 5,4 3,5

3
 • Solution concept?

— Does A have a dominant strategy? No. Prefers S if B chooses S,

otherwise prefers L.
— Does B have a dominant strategy? Yes. Prefers L regardless of A0 s
choice.
— If A and B are both rational, then A should correctly anticipate that
B will choose L.
— Accordingly, A should choose L.
— Equilibrium is L, L.
— This is a ‘pure strategy’ equilibrium in that players just choose one
action, do not randomize to ‘mix’ strategies.

• How do we know this is an equilibrium?

— Player B has no incentive to deviate given A0 s choices.

— Player A has no incentive to deviate given B 0 s choices.
— Hence, these strategies are complementary, or self-enforcing.

• This brings us to first core concept in game theory: Nash Equilibrium.

1.2 Nash Equilibrium

Definition 1 A pair of strategies (a∗ , b∗ ) (for a 2-player game) is a called a
Nash Equilibrium iff:
a∗ is player A0 s best strategy when B is playing b∗ and b∗ is player B 0 s best
strategy when A is playing a∗ .

• Truthful revelation: At a Nash equilibrium, one player would not benefit

from knowing the strategy of the other in advance since A would never
have incentive to play anything other than a∗ if B is playing b∗ .
• (This is actually quite similar to the equilibrium concept in the signaling
and adverse selection models we just completed. Expectations of buy-
ers/sellers or workers/employers in those models must be complementary.
In equilibrium, everyone’s expectation of quality and price, or productivity
and wages are mutually consistent.)
• This concept seems very obvious. But it was not formalized until Nash
wrote it down and proved that every game has at least one Nash equilib-
rium.
• Not all games have a pure strategy equilibrium, however.

4
1.3 Example: Rock, paper, scissors
• Recall the childhood game of rock paper scissors. This is a game where you
hold out one of three symbols corresponding to rock, paper and scissors.
The rules are:

— Paper covers rock

— Rock breaks scissors
— Scissors cuts paper
— (Any 2 identical objects are a tie; begin again)

• So, the payoff matrix looks like

B
Rock Paper Scissors
Rock 0, 0 −1, 1 1, −1
A Paper 1, −1 0, 0 −1, 1
Scissors −1, 1 1, −1 0, 0
• There is no pure strategy equilibrium to this game. If player A announced
a pure strategy (e.g., play rock), player B would choose a pure strategy
that would defeat A, and then A would have incentive to deviate from the
original strategy.
• But there is an obvious mixed strategy—that most children understand
intuitively.

1.4 Another mixed strategy example

• The “Family Vacation” game. Family goes on vacation together. Parents
prefer Paris, kids prefer Disney. But neither prefers to go without the
other (that’s why it’s a family vacation).
• Payoffs:
Kids
Disney Paris
Parents Disney 1, 2 0, 0
Paris 0, 0 2, 1
• Clearly, this game has two pure strategy Nash equilibria: D, D and P, P .
If kids were certain that parents were playing P , then they would play P
too, and vice versa if parents were certain kids were playing D.
• This game also has a mixed strategy equilibrium. What is it?
• Say parents play Paris with probability P , play Disney with probability
1 − P.
• Say kids play Paris with probability K , play Disney with probability 1−K.

5
• What is the expected utility of the parents?

E[Up ] = 2 · P K + 0 · (1 − P )K + 0 · K(1 − P ) + 1 · (1 − P )(1 − K)

= 3KP − P − K + 1

• Notice that E[Up ] depends on K.

— If K = 13 , then E[Up ] = 23 , which is independent of P.

— If K > 13 , then E[Up ] is maximized by setting P = 1, always going
to Paris.
— If K < 13 , then E[Up ] is maximized by setting P = 0, always going
to Disney.

• What is the expected utility of the kids?

E[Uk ] = 2 · (1 − P )(1 − K) + 0 · (1 − P )K + 0 · K(1 − P ) + 1 · KP

= 3KP − 2P − 2K + 2

• So E[Uk ] depends on P.

— If P = 23 , then E[Uk ] = 23 , which is independent of K.

— If P > 23 , then E[Uk ] is maximized by setting K = 1, always going
to Paris.
— If P < 23 , then E[Uk ] is maximized by setting K = 1, always going
to Disney.

• Hence, a mixed strategy equilibrium for this game has (K = 13 , P = 23 ),

and E(Up ) = E(Uk ) = 23 .
• Notice by the way the frequencies that result:
Kids
Disney
¡ ¢ Paris
¡ ¢ .
Parents Disney 1, 2 ¡ 29 ¢ 0, 0 ¡ 19 ¢
Paris 0, 0 49 2, 1 29
5
• So, in 9 of cases, both kids and parents get 0, which is not desirable.
• The problem: this is a coordination game—where both players benefit by
playing complementary strategies. By randomizing strategies, the non-
coordinated outcomes occurs most often (more than 1/2 the time).
• Nash equilibrium should not be confused with a competitive market equi-
librium which, under conditions that we are familiar with, has desirable
efficiency properties.
• In game theory, as in information economics, equilibria result from strate-
gic interactions rather than atomistic maximizing behavior. And so there
is no reason to think that these outcomes will be desirable.

6
1.5 Prisoner’s dilemma
• You are already familiar with the prisoner’s dilemma. Review it briefly
anyway because it underscores a central issue in non-cooperative games —
the issue of credible commitment.
• The setup: Two criminals accused of a crime. The district attorney pulls
each aside to say: if you help me convict (‘rat on’) the other prisoner, I’ll
let you go free and the other prisoner will get 10 years. However, if you
both rat on one another, you’ll each get 5 years. The prisoners understand
that if neither rats on the other, they will each serve only 3 years. So, the
payoff matrix:
B
Rat Not
A Rat −5, −5 0, −10
Not −10, 0 −3, −3
• As is clear, Rat, Rat is the only Nash equilibrium of this game, even
though both prisoners would be strictly better off if they could choose
N ot Rat, N ot Rat.
• Hence, the prisoner’s dilemma underscores one of the interesting proper-
ties of game theoretic models: Outcomes that appear optimal are often
not stable when subject to the Nash criterion. In fact, they are often
dominated.
• This raises a set of questions:

1. When are commitments credible?

2. When should threats be believed?
3. Is there anyway to ensure cooperation?

• These issues are explored a bit in the tragedy of the commons game

1.6 Tragedy of the commons

• Two yak herders, A and B.
• There is one common land that can only support so many yaks, who graze
and produce milk. If the common is overgrazed, total milk production will
fall.
• Milk sells for $1 per gallon. Let V equal milk per yak, and YA , YB equal the
number of yaks brought to the common by each herder. The production
function is:
V (YA , YB ) = 200 − (YA + YB )2

7
 2
• Notice ∂V ∂ V
∂Y < 0, ∂Y 2 < 0, meaning that each Yak does increasing marginal
damage (note, the outcome variable is in gallons per yak, so obviously it
will be optimal to bring some positive number of yaks).
A 21#2

Graze 4 Graze 5

B B

G4 G6 G4 G6
G5 G5

(544,544) (476,595) (400,600) (545,476) (500,500) (395,474)

• Solution concept: What does each Yak herder do taking the actions of the
other herder as given?
• Problem for herder A

max YA · V = 200YA − YA (YA + YB )2

YA
∂
= 200 − (YA + YB )2 − 2YA (YA + YB )
∂YA
F OC : 200 − 3YA2 − YB2 − 4YA YB = 0

• Problem for herder B is symmetric:

F OC : 200 − 3YB2 − YA2 − 4YA YB = 0

• Given symmetry, it must be YA = YB , implying that:

200 − 3Y 2 − Y 2 − 4Y 2 = 0,
200 = 8Y 2 ,
YA = YB = 5.

• Each herder brings 5 yaks to the common and earns 5 · (200 − 102 ) = 500,
and total output is 1, 000 gallons of milk.

8
• What would be the social optimum?

max Y · V = 200Y − Y 3 ,
Y
∂
= 200 − Y · Y 2 = 0,
∂Y
10 √
Y = 6 ≈ 8.16,
3
and total output is

8.16 · 200 − 8.163 = 1, 089.

• So clearly, the common is being overgrazed. The Nash equilibrium is not

a social optimum.
• But what if one herder promised to only bring 4 yaks? So an idea: players
could announce their strategies in advance.
• Focus on the game in extensive form where A is trying to decide whether to
graze 4 or 5 yaks. Prior to A0 s move, B will announce to A the strategies
he will play (to graze 4, 5 or 6 yaks) in response to A0 s choice. So, the
setup:

1. B announces the action he will take if A brings 4 yaks and the action
he will take if A brings 5 yaks.
2. A takes his action.
3. B takes his action (hence, this is now a sequential game).

B
4, 4 4, 5 4, 6 5, 4 5, 5 5, 6 6, 4 6, 5 6, 6
4 544, 544 544, 544 544, 544 476, 595 476, 595 476, 595 400, 600 400, 600 400, 600
5 595, 476 500, 500 395, 474 595, 476 500, 500 395, 474 595, 476 500, 500 385, 474

• What are the pure strategy Nash equilibria of this game?

• Looking at the payoffs, A will choose his best response to each of B 0 s
strategies. These are

B
4, 4 4, 5 4, 6 5, 4 5, 5 5, 6 6, 4 6, 5 6, 6
4 544, 544 544, 544 476, 595 400, 600
5 595, 476 595, 476 500, 500 595, 476 500, 500

• If B assumes A is rational, he anticipates these choices and so the following

would be Nash equilibria:

1. B threatens 6, 6; A chooses 4: Payoffs are 400, 600

9
 2. B threatens 5, 5; A chooses 5: Payoffs are 500, 500
3. B threatens 6, 5; A chooses 5: Payoffs are 500, 500

B
4, 4 4, 5 4, 6 5, 4 5, 5 5, 6 6, 4 6, 5 6, 6
4 400, 600
5 500, 500 500, 500

• Hence, this game has three Nash equilibria. Question: Are any of these
equilibria problematic?
• Yes, both 1 and 2 involve non-credible threats. Player B threatens to take
actions contingent’s on A0 s choice that it would not be rational for him
to take. For example, if A brought 5 Yaks, then B should also bring 5.
Player B should never bring 6 Yaks if player A brings more than 4.
• So the problem is that B is making threats that should not be believed.
But in a simple Nash equilibrium, they are believed. The Nash equilbrium
concept does not rule out these implausible threats. Why not?
• Because these threats don’t have to be carried out in equilibrium. If A
brings 4 yaks, then B brings 6, which is rational. And if A brings 5, then
B brings 5, which is also rational. So player B is never forced to carry out
an irrational threat, which would in fact violate the Nash equilibrium.
• This example points to a problem with the Nash concept, which is that
implausible beliefs about what would happen ‘off the equilibrium path,’
can lead to implausible results in equilibrium (e.g., Player A brings 4 yaks
in response to B 0 s threat to bring 6).
• This motivates the ‘equilibrium refinement’ of Subgame Perfection.

Definition 2 Subgame: Stage of the game at t > 1 moves.

Definition 3 Subgame perfection: A Nash equlibrium is subgame perfect if the

players’ strategies constitute a Nash equilibrium in every subgame.

• Translation: a Nash equilibrium is sub-game perfect if it does not require

a non-credible threat to be sustained.
• B 0 s threat to bring 6 yaks if A brought 5 failed the subgame perfection
test because if the players reached this subgame—i.e., player A did bring 5
yaks—player B would necessarily bring 5 yaks.

10
1.7 Subgame perfection: The Scorsese example
• In the (poor) 1996 film “Casino,” Joe Pesci plays a gangster who is sent to
collect $50, 000 from a banker. The banker has the money in a safe. Pesci
threatens the banker with a baseball bat. The problem is that only the
banker knows the code to the safe. If Pesci hits the banker with the bat,
he probably won’t get the money. But he will go to jail. The banker gives
him a knowing look... So, the payoffs look as follows (note— the Banker
moves first):

Pesci
Assault Not
Banker Give Money −50, 000, +50, 000+Jail −50, 000, +50, 000
No Money 0+Pain,Jail 0, 0

Banker
21#3

Give $50,000 Not

Pesci Pesci

Assault Not Assault Not

-50K, +50K -50K, +50K Pain, Jail 0,0

+ Pain + Jail

Banker moves first

• The banker recognizes that it is not subgame perfect for Pesci to assault
him since Pesci only has jail time to gain.
• So, the subgame perfect equilibrium of this game is that the banker keeps
the money and Pesci leaves without committing the assault.
• But this is not what happens... Pesci convinces the banker he is irrational
and gets the money.
• It’s a great scene—and a nice application of game theory.

11
1.8 Related: Changing your payoffs
• Subgame perfection demonstrates that non-credible threats should not be
believed. This works to the disadvantage of the person making the threats.
There are at least two ways to change the game.
• The Scorsese example demonstrates one way: convince your opponent that
you are irrational. Irrational people may carry out threats that are self-
destructive.
• [North Korea is probably the master of this strategy.]
• A second strategy to make threats credible is, somewhat paradoxically, to
make your payoffs worse—that is, destroy your fallback option.

1. Schelling island example.

2. Quitting drinking? If you could stop drinking for a while, you would
no longer be addicted. But the craving is so intense after 24 hours,
that you cannot credibly commit to stop. Take mucilage—does severe
liver damage if consumed with alcohol.

1.9 The problem of cooperation

• Returning to the problem underscored by the prisoner’s dilemma, game
theory seems to predict very little cooperation in the world. For example,
consider the centipede game.
• A sequential move game: 1st player 1, then player 2, then player 1, etc.

Sequential move game 21#4

1 C 2C 1 C 2 C 1 C 2 1 C 2 C 1 C 2C
100,
100
D D D D D D D D D D

1 0 2 1 3 2 98 97 99 98
1 3 2 4 3 5 98 100 99 101

• Both players benefit from cooperation.

• But there is always a temptation to defect.

12
 • This game is solved by backward induction. Start at the last node, work
your way back. As you can see, cooperation is clearly dominated in the
final play. But then it is dominated in the 2nd to last move. And so on...
• Iterated dominance leaves only one equilibrium, and it is an undesirable
one.
• This particular prediction does not have the ring of truth. Puts a hefty
premium on the rationality of both players, and their belief in the ratio-
nality of one another. It is also not especially well supported by empirical
evidence
• What would you do if the 1st player did not defect on move one? Game
theory does not make a clear prediction. This is ‘off the equilibrium path.’
• The problem of cooperation appears vexed.

1.10 Idea: Repeated interactions

• One notion is that if players repeatedly face the same situation, they may
recognize that it is in their mutual interest to cooperate. So, maximize
the pie rather than simply grabbing the largest share of a small pie.
• Consider a three period repeated prisoner’s dilemma. Like a standard
prisoner’s dilemma except that if both players cooperate, there is a con-
tinuation game, up to a total of 3 periods. Assume also that there is a
discount factor of δ < 1.
• Period 1
B
Coop Defect
A Coop 5, 5 −1, 10
Defect 10, −1 1, 1

• If they cooperate in period 1, then period 2 looks similar. So does period

3. Then the game ends.

B
Coop Defect
A Coop 5, 5 −1, 10
Defect 10, −1 1, 1

• So the payoff to cooperation is:

(5, 5) + (5δ, 5δ) + (5δ 2 , 5δ 2 ),

whereas the payoff to defection is

(1, 1) + (δ, δ) + (δ 2 , δ 2 ).

13
 • Obviously, it is advantageous to cooperate.
• But of course in period 3, each ought to defect, since there is no further
payoff to cooperation. And so by backward induction, each should defect
in 2, and then in 1. Hence, no honor among thieves.

1.11 Infinitely repeated games

• What causes games to unravel? It appears to be related to backward
induction. Because ‘defection’ is optimal in the last period, it becomes
optimal in the second to last period, then the third to last, etc. What if
there was no last period?
• Examples:

— Are you more likely to tip at a local restaurant that you go to often
or at an out-of-town restaurant where you’ll likely never return?
— Many drivers are extremely rude in city traffic. But they would not do
this if those drivers were their neighbors (even annoying neighbors).
— “No one in the history of the world has ever washed a rented car.” —
Lawrence Summers, Economist, President of Harvard University.

• In an infinitely repeated game (or a game with no known end point), there
cannot be backward induction. No player knows when the last period will
occur. So, the unravelling that we’ve seen above may not occur.
• Consider the following strategy: Each player announces she will cooperate
so long as the other player does so. But if the one player defects, the other
will punish her with no further cooperation. This is called the “grim
trigger strategy.” Is it a Nash equilibrium?
• Payoff to cooperation is
5
5 + 5δ + 5δ 2 + 5δ 3 + ... + 5δ ∞ = .
1−δ
Payoff to defection is
δ
10 + δ + δ 2 + ... + δ ∞ = 10 + .
1−δ
• Cooperation is therefore an equilibrium iff:
5 δ
> 10 + ,
1−δ 1−δ
5−δ
> 10,
1−δ
5 > 10 − 9δ
5
δ > ≈ .556
9

14
• So, if the future is sufficiently important, cooperation is sustainable as an
equilibrium.
• Hence, there can be ‘cooperative’ outcomes in non-cooperative game the-
ory, but seemingly only under restrictive conditions.
• Are the predictions of this models too strong?

15